In order to understand the process which leads to the extraction of the judgment outcome, the general structure of a Swiss court ruling is briefly introduced. A typical Swiss verdict is comprised of five sections: the header, the facts, the considerations, the ruling and the footer. Note that not all case rulings contain all listed sections below. Some sections may or may not be absent.
- The header represents the introduction to the respective judgement. It names the court and its composition, as well as the place and date of the decision. It introduces the parties and their respresentatives and may briefly touch upon the matter of dispute, without providing a detailed explanation.
- The facts serve to present the disputed and undisputed facts, the request as well as the factual assertion of the parties, the history of the case and any taking of evidence.
- In the considerations, the court makes a legal assessment of the entire events. This constitutes the legal reasoning that later leads to the decision.
- The ruling contains the decision on the merits and the costs.
- The footer usually contains some information on the legal remedies. Often times an official footer section is omitted and the contents are added to the end of the ruling section.
The required steps leading to the extraction of the judgment outcome for a given case document can be described in six steps. The order of tasks usually follow the numeration depicted in the figure below, yet can vary as every step can be executed in isolation.
Entscheidsuche.ch updates their website daily with all publicly available caseses from the courts in Switzerland. Per default, only case documents which do not already exist locally are considered in order to avoid rewriting the entire database. In a first step, all the files are scraped including the associated metadata of every court's folder from entscheidsuche.ch/docs. A combination of BeautifulSoup's parsing library and Requests is used to extract the complete list of URLs to each of the case documents. The list is then iterated to download all the files.
The module can be executed by running:
python -m scrc.preprocessors.scraper
Since within a Swiss court case documents can vary in the language they are written in, every case must have assigned an appropriate language identifier which will be needed for the section splitting and judgment outcome extraction task. For every document within our database for which we were unable to extract the corresponding language from the metadata, we make use of the fastText language identification tool to assign to each the appropriate language id.
The module can be executed by running:
python -m scrc.preprocessors.language_identifier
Certain courts may include strange patterns or unnecessary text within their case documents which can make the entire process more prone to errors. To improve further processing, cleaning functions for certain courts have been implemented removing such patterns and texts. The cleaners can be found in the cleaning_regexes.json file.
The module can be executed by running:
python -m scrc.preprocessors.extractors.cleaner
Make sure to remove the desired courts from the list in spiders_cleaned.txt
The section splitting task serves the purpose of splitting the case ruling in its predefined sections. It is done so by defining a set of regex indicators, each of which implicate the start of one of the sections. The composition of words and special characters to indicate certain sections however, may differ for each court in Switzerland. Thus for every Swiss court, a different set of regex indicators has to be defined for each section which may appear within a ruling.
Given a case document, it is first broken down into its paragraphs. Using the set of regex indicators we iterate through each paragraph and search for a match. Once a match has been found, all subsequent paragraphs are appended to the specific section until we find a new match. We continue this process until we reach the end of the file.
The module can be executed by running:
python -m scrc.preprocessors.extractors.section_splitter
Make sure to remove the desired courts from the list in spiders_section_split.txt
The main issue facing the section splitting task lies in the wide variety of words used to indicate the start of each section and the different variations and additives these words may encompass. Not only may courts amongst themselves use different variations but this can also occur within the same court. Because of this, a manual extraction of indicators for a single court may require hours of structural analysis of different case rulings within that court. In order to significantly reduce manual analysis while addressing maintainability for future courts and cases, a python module has been implemented which presents common indicators for a specific court for each section. It works by counting the occurrence of every paragraph appearing across all case rulings for a given court. If a court has certain patterns of indicating different sections in certain ways, those will be reflected by having a significantly higher occurrence count than a random paragraph appearing within a case ruling. Using this logic we can then filter out low occurrence paragraphs and make rough assignment choices by having a baseline of predetermined common indicators for each section. The result is an accurate overview of all appearing patterns of indicators for each of the sections.
The pattern_extractor.py module can also be used to output the coverage of courts and their sections by executing the module with an argument.
The module can be executed by running:
python -m scrc.preprocessors.extractors.pattern_extractor | python -m scrc.preprocessors.extractors.pattern_extractor 1 (to output only the section coverage in the terminal)
Make sure to remove the desired courts from the list in pattern_extraction.txt
Top four entries for consideration section of Verwaltungsgericht Bern
| total count | indicator | coverage (%) |
|---|---|---|
| 7160 | Erwägungen: | 95.51 |
| 125 | Der Einzelrichter zieht in Erwägung: | 1.67 |
| 94 | Sachverhalt und Erwägungen:_ | 1.25 |
| 56 | Der Einzelrichter zieht in Erwägung, | 0.75 |
The judgment outcome extraction step serves the purpose of extracting the judgment outcome based on a set of predefined indicators. We have defined seven possible outcomes so far:
- Approval: The request is accepted in full
- Partial approval: Part of the request is accepted
- Dismissal: The request is dismissed in its entirety
- Partial dismissal: Part of the request is rejected
- Inadmissible: The court does not have jurisdiction over the request; Formal defects within the complaint
- Write off: No reason for the proceedings (the decision is no longer needed, for example, because the parties have reached an out-of-court settlement or because two proceedings have been combined).
- Unification: When two cases are about the same thing, they will get merged into one.
A helper method similar to the section splitting can be used for the extraction of indicators in the judgment outcome task. Instead of looking at whole paragraphs however, the ruling section of each case will be split into different set of n-grams and compared to each other. Patterns which tend to remain consistent across many cases within a court will have a significantly higher occurrence count. These consistencies will primarily be the indicators responsible for the judgment outcome.
The judgment_pattern_extractor.py module can also be used to output the coverage of courts and their judgment outcome by executing the module with an argument.
The module can be executed by running:
python -m scrc.preprocessors.extractors.judgment_pattern_extractor | python -m scrc.preprocessors.extractors.judgment_pattern_extractor 1 (to output only the judgment coverage in the terminal)
Make sure to remove the desired courts from the list in judgment_pattern_extractor.txt
Top six entries for a combination of three neighboring words for the cantonal court of Vaud.
| total count | indicator |
|---|---|
| 22145 | ('Le', 'recours', 'est') |
| 12292 | ('recours', 'est', 'rejeté.') |
| 4238 | ('recours', 'est', 'admis.') |
| 1971 | ('recours', 'est', 'irrecevable.') |
| 1493 | ('recours', 'est', 'partiellement') |
| 1422 | ('est', 'partiellement', 'admis.') |
The coverage_verification.py module can be used to output x amount of cases with their respective sections and judgment outcomes as a docx file to quickly scim through the decisions and verify them on their correctness. Refer to the module for further information.
The module can be executed by running:
python -m scrc.analyses.coverage_verification | python -m scrc.analyses.coverage_verification 1 (to output only the ruling section with the judgment)
Make sure to remove the desired courts from the list in coverage_verification.txt